Accomplished since last meeting¶
- Another 1900 addresses were able to be geocoded
- Relative coordinates were converted to x,y pairs and a model ready data set was created
- Progressive time models completed
- High risk kids (9 boys and 7 girls) from the sample models were extracted for to get their growth curves
- Sample figures for the paper are created
- A function to create table outputs for summary statistics of key features is created
- Acuracy comparisons for the best models compared to Mary Jo's original and updated inclusion criteria
BMI Percentile Distribution for All Children in the Cohort¶
In [4]:
fig, ax = plt.subplots(figsize=(12,8))
index = np.arange(len(bins))
bar_width = 0.4
opacity = 0.8
bar1 = plt.bar(index, boys_bins, bar_width, alpha=opacity,
label='Boys (n={0:,.0f}; $\mu$={1:,.2f})'.format(np.sum(boys_bins), np.mean(boys)))
bar2 = plt.bar(index + bar_width, girls_bins, bar_width, alpha=opacity,
label='Girls (n={0:,.0f}; $\mu$={1:,.2f})'.format(np.sum(girls_bins), np.mean(girls)))
# plt.xlabel('Obesity Percentiles', fontsize=14)
plt.ylabel('Number of Children', fontsize=14)
plt.title('BMI Percentile Distribution at 4.5 to 5.5 years - No Exclusions', fontsize=20)
# plt.xticks(index + bar_width, ['{0:.1f}%'.format(b * 100) for b in bins])
plt.xticks(index, bin_names, rotation=30)
ax.yaxis.grid(linestyle='--')
ax.set_axisbelow(True)
plt.legend(fontsize=12)
plt.tight_layout()
# plt.savefig('../outputs_age_analyses20180221/no_exclusions_bmi_percentiles.png', dpi=96)
plt.show()
Growth Curves for the 16 Highest Risk Kids in the Data Set¶
In [7]:
train.plot_growth_curve(d1,None,keys[0], hide_mrn=True)
In [8]:
train.plot_growth_curve(d1,None,keys[1], hide_mrn=True)
In [9]:
train.plot_growth_curve(d1,None,keys[2], hide_mrn=True)
In [10]:
train.plot_growth_curve(d1,None,keys[3], hide_mrn=True)
In [11]:
train.plot_growth_curve(d1,None,keys[4], hide_mrn=True)
In [12]:
train.plot_growth_curve(d1,None,keys[5], hide_mrn=True)
In [13]:
train.plot_growth_curve(d1,None,keys[6], hide_mrn=True)
In [14]:
train.plot_growth_curve(d1,None,keys[7], hide_mrn=True)
In [15]:
train.plot_growth_curve(d1,None,keys[8], hide_mrn=True)
In [16]:
train.plot_growth_curve(d1,None,keys[9], hide_mrn=True)
In [17]:
train.plot_growth_curve(d1,None,keys[10], hide_mrn=True)
In [18]:
train.plot_growth_curve(d1,None,keys[11], hide_mrn=True)
In [19]:
train.plot_growth_curve(d1,None,keys[12], hide_mrn=True)
In [20]:
train.plot_growth_curve(d1,None,keys[13], hide_mrn=True)
In [21]:
train.plot_growth_curve(d1,None,keys[14], hide_mrn=True)
In [22]:
train.plot_growth_curve(d1,None,keys[15], hide_mrn=True)
Results for the most predictive feature: 'Vital: BMI-latest'¶
In [24]:
modelix = 'BMI'
plt.figure(figsize=(9,9))
for ix in range(len(prec_total)):
if modelix not in titles_total[ix] or 'randomforest' in titles_total[ix] or 'gradientboost' in titles_total[ix]:
continue
if 'girls' in titles_total[ix]:
plt.plot(1- np.array(spec_total[ix]), np.array(recall_total[ix]), linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
plt.plot(1- np.array(spec_total[ix]), np.array(recall_total[ix]), linestyle='--', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
plt.legend(fontsize = 10)
plt.xlabel('1 - Specificity', fontsize=12)
plt.ylabel('Sensitivity', fontsize=12)
plt.axis('equal')
plt.title('ROC Curve: Obesity Predicted at 5 years - "Vital: BMI-latest"', fontsize=14)
plt.grid(True)
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_ROC.png', dpi=300)
plt.show()
In [25]:
plt.figure(figsize=(9,9))
for ix in range(len(prec_total)):
if modelix not in titles_total[ix] or 'randomforest' in titles_total[ix] or 'gradientboost' in titles_total[ix]:
continue
if 'girls' in titles_total[ix]:
plt.plot(recall_total[ix], prec_total[ix], linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
plt.plot(recall_total[ix], prec_total[ix], linestyle='--', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
plt.xlabel('Recall (Sensitivity)', fontsize=14)
plt.ylabel('Precision (PPV)', fontsize=14)
plt.legend(fontsize = 8)
plt.axis('equal')
plt.title('Precision-Recall Curve: Obesity Predicted at 5 years - "Vital: BMI-latest"', fontsize=14)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
LASSO and Random Forest Comparison for Boys and Girls¶
In [26]:
plt.figure(figsize=(9,9))
for ix in range(len(prec_total)):
if any(x in titles_total[ix] for x in ('w/o vitals','Wt','no_maternal','w/o exclusions', 'BMI','gradientboost')):
continue
if 'girls' in titles_total[ix]:
plt.plot(1- np.array(spec_total[ix]), np.array(recall_total[ix]), linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
plt.plot(1- np.array(spec_total[ix]), np.array(recall_total[ix]), linestyle='--', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
plt.legend(fontsize = 9)
plt.xlabel('1 - Specificity', fontsize = 14)
plt.ylabel('Sensitivity', fontsize = 14)
plt.axis('equal')
plt.title('ROC Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid(True)
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_ROC.png', dpi=300)
plt.show()
In [27]:
plt.figure(figsize=(9,9))
for ix in range(len(prec_total)):
if any(x in titles_total[ix] for x in ('w/o vitals','Wt','no_maternal','w/o exclusions', 'BMI','gradientboost')):
continue
if 'girls' in titles_total[ix]:
plt.plot(recall_total[ix], prec_total[ix], linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
plt.plot(recall_total[ix], prec_total[ix], linestyle='--', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
plt.xlabel('Recall (Sensitivity)', fontsize = 14)
plt.ylabel('Precision (PPV)', fontsize = 14)
plt.legend(fontsize = 8)
plt.axis('equal')
plt.title('Precision-Recall Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
In [28]:
plt.figure(figsize=(18,9))
plt.subplot(1, 2, 1)
for ix in range(len(prec_total)):
if any(x in titles_total[ix] for x in ('w/o vitals','Wt','no_maternal','w/o exclusions', 'BMI','gradientboost')):
continue
if 'girls' in titles_total[ix]:
plt.plot(recall_total[ix], prec_total[ix], linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
continue
plt.ylabel('Precision (PPV)', fontsize = 14)
plt.legend(fontsize = 8, loc=8)
plt.axis('equal')
plt.title('Girls Precision-Recall Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid()
plt.tight_layout()
plt.subplot(1, 2, 2)
for ix in range(len(prec_total)):
if any(x in titles_total[ix] for x in ('w/o vitals','Wt','no_maternal','w/o exclusions', 'BMI','gradientboost')):
continue
if 'boys' in titles_total[ix]:
plt.plot(recall_total[ix], prec_total[ix], linestyle='-', label=titles_total[ix]+' - AUC={:0.2f}'.format(auc_list[ix][0]))
else:
continue
plt.legend(fontsize = 8, loc=8)
plt.axis('equal')
plt.xlabel('Recall (Sensitivity)', fontsize = 14)
plt.title('Boys Precision-Recall Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
Results for the best model each prediction age¶
In [30]:
plt.figure(figsize=(9,9))
for l in top_ix:
if 'girls' in titles_total[l[2]]:
plt.plot(1- np.array(spec_total[l[2]]), np.array(recall_total[l[2]]), linestyle='-', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
else:
plt.plot(1- np.array(spec_total[l[2]]), np.array(recall_total[l[2]]), linestyle='--', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
plt.legend(fontsize = 10)
plt.xlabel('1 - Specificity', fontsize=14)
plt.ylabel('Sensitivity', fontsize=14)
plt.axis('equal')
plt.title('ROC Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid(True)
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_ROC.png', dpi=300)
plt.show()
In [31]:
plt.figure(figsize=(9,9))
for l in top_ix:
if 'girls' in titles_total[l[2]]:
plt.plot(recall_total[l[2]], prec_total[l[2]], linestyle='-', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
else:
plt.plot(recall_total[l[2]], prec_total[l[2]], linestyle='--', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
plt.xlabel('Recall (Sensitivity)', fontsize=14)
plt.ylabel('Precision (PPV)', fontsize=14)
plt.legend(fontsize = 10)
plt.axis('equal')
plt.title('Precision-Recall Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
In [32]:
plt.figure(figsize=(20,10))
plt.subplot(1, 2, 1)
for l in top_ix:
if 'girls' in titles_total[l[2]]:
plt.plot(recall_total[l[2]], prec_total[l[2]], linestyle='-', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
else:
continue
plt.ylabel('Precision (PPV)', fontsize = 14)
plt.legend(fontsize=10, loc=8)
plt.axis('equal')
plt.title('Girls Precision-Recall Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid()
plt.tight_layout()
plt.subplot(1, 2, 2)
for l in top_ix:
if 'boys' in titles_total[l[2]]:
plt.plot(recall_total[l[2]], prec_total[l[2]], linestyle='-', label=titles_total[l[2]]+' - AUC={:0.2f}'.format(auc_list[l[2]][0]))
else:
continue
plt.legend(fontsize=10, loc=8)
plt.axis('equal')
plt.xlabel('Recall (Sensitivity)', fontsize = 14)
plt.title('Boys Precision-Recall Curve: Obesity Predicted at 5 years', fontsize = 18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
All Top Models¶
In [35]:
df_l = []
headers=['ix','title','auc','auc ste']
for ix,(auc,title) in enumerate(zip(auc_list,titles_total)):
df_l.append([ix,title,auc[0],auc[1]])
df = pd.DataFrame(df_l, columns=headers)
top_ix = []
for age in ['@ 6','@ 12','@ 18','@ 24','@ 36','@ 48']:
for g in ['boys','girls']:
filtered = df[(df.title.str.contains(age)) &(df.title.str.contains(g)) & (df.title.str.contains('w/o exclusions')) & (~df.title.str.contains('no_maternal'))].sort_values(by='auc', axis=0, ascending=False)
print(filtered.head(3))
name = filtered.iloc[0,1].replace(" maternal w/o exclusions - model: ", " ")
top_ix.append([age,g,filtered.iloc[0,0],name])
Comparison of Mary Jo's Inclusion Criteria to Pediatric Study¶
- Using the best model for boys and girsl where prediction is at 24 months
- Original: base parameters for the pediatric study where no inclusion criteria is used
- Inclusion: uses Mary Jo's inclusion criteria
- Age of mother at birth is at least 18 years old
- No complications with the child delivery
- Mother is Hispanic/Latina
- Modified Inclusion: the same as above but all ethnicities are allowed
In [40]:
plt.figure(figsize=(9,9))
for i,title in enumerate(titles_24):
if 'original' in titles_24[i].lower():
plt.plot(1- np.array(spec_24[i]), np.array(recall_24[i]), linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
elif 'inclusion mod' in titles_24[i].lower():
plt.plot(1- np.array(spec_24[i]), np.array(recall_24[i]), linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
else:
plt.plot(1- np.array(spec_24[i]), np.array(recall_24[i]), linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
plt.legend(fontsize = 10)
plt.xlabel('1 - Specificity', fontsize=14)
plt.ylabel('Sensitivity', fontsize=14)
plt.axis('equal')
plt.title('ROC Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid(True)
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_ROC.png', dpi=300)
plt.show()
In [43]:
plt.figure(figsize=(9,9))
for i,title in enumerate(titles_24):
if 'original' in titles_24[i].lower():
plt.plot(recall_24[i], prec_24[i], linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
elif 'inclusion mod' in titles_24[i].lower():
plt.plot(recall_24[i], prec_24[i], linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
else:
plt.plot(recall_24[i], prec_24[i], linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_24[i]))
plt.xlabel('Recall (Sensitivity)', fontsize=14)
plt.ylabel('Precision (PPV)', fontsize=14)
plt.legend(fontsize = 10)
plt.axis('equal')
plt.title('Precision-Recall Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
Comparison of Mary Jo's Inclusion Criteria to Pediatric Study¶
- Using the best model for each prediction age for boys and girls
- Original: base parameters for the pediatric study where no inclusion criteria is used
- Inclusion: uses Mary Jo's inclusion criteria
- Age of mother at birth is at least 18 years old
- No complications with the child delivery
- Mother is Hispanic/Latina
- Modified Inclusion: the same as above but all ethnicities are allowed
In [47]:
plt.figure(figsize=(20,10))
plt.subplot(1, 2, 1)
for i,title in enumerate(titles_best):
if 'girls' in titles_best[i]:
if 'original' in titles_best[i].lower():
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
elif 'mod.' in titles_best[i].lower():
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
continue
plt.legend(fontsize = 10)
plt.ylabel('Sensitivity', fontsize=14)
plt.xlabel('1 - Specificity', fontsize=14)
plt.axis('equal')
plt.title('Girls ROC Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid(True)
plt.tight_layout()
plt.subplot(1, 2, 2)
for i,title in enumerate(titles_best):
if 'boys' in titles_total[i]:
if 'original' in titles_best[i].lower():
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
elif 'mod.' in titles_best[i].lower():
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
plt.plot(1- np.array(spec_best[i]), np.array(recall_best[i]), linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
continue
plt.legend(fontsize = 10)
plt.xlabel('1 - Specificity', fontsize=14)
plt.axis('equal')
plt.title('Boys ROC Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid(True)
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
In [49]:
plt.figure(figsize=(20,10))
plt.subplot(1, 2, 1)
for i,title in enumerate(titles_best):
if 'girls' in titles_best[i]:
if 'original' in titles_best[i].lower():
plt.plot(recall_best[i], prec_best[i], linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
elif 'mod.' in titles_best[i].lower():
plt.plot(recall_best[i], prec_best[i], linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
plt.plot(recall_best[i], prec_best[i], linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
continue
plt.ylabel('Precision (PPV)', fontsize=14)
plt.xlabel('Recall (Sensitivity)', fontsize=14)
plt.legend(fontsize = 8)
plt.axis('equal')
plt.title('Girls Precision-Recall Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid()
plt.tight_layout()
plt.subplot(1, 2, 2)
for i,title in enumerate(titles_best):
if 'boys' in titles_best[i]:
if 'original' in titles_best[i].lower():
plt.plot(recall_best[i], prec_best[i], linestyle='-', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
elif 'mod.' in titles_best[i].lower():
plt.plot(recall_best[i], prec_best[i], linestyle=':', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
plt.plot(recall_best[i], prec_best[i], linestyle='--', label=title+' - AUC={:0.2f}'.format(auc_best[i]))
else:
continue
plt.xlabel('Recall (Sensitivity)', fontsize=14)
plt.legend(fontsize = 8)
plt.axis('equal')
plt.title('Boys Precision-Recall Curve: Obesity Predicted at 5 years', fontsize=18)
plt.grid()
plt.tight_layout()
# plt.savefig(newdir+'/Pediatric_Girls_PR.png', dpi=300)
plt.show()
Next Steps¶
- Rerun analyses at each of the prediction time points with:
- all data
- all data excluding weight and BMI readings
- all data excluding census data
- predict for extreme obesity (99%?)
- Update address data from the most recent data release
- Will need to be geocoded
- Census features for at birth, 2 years old (or month we are predicting from), 1 year before birth?
- Report back on results with cleaned up plots